IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES 1 State Variable-Based Transient Analysis Using Convolution Carlos E. Christoffersen, Student Member, IEEE, Mete Ozkar, Student Member, IEEE, Michael B. Steer Fellow, IEEE, Michael G. Case, Member, IEEE, and Mark Rodwell, Member, IEEE Abstract— A state variable-based approach to the impulse response and convolution analysis of distributed microwave circuits is developed. The state-variable approach minimizes computation time and memory requirements. It allows the use of parameterized nonlinear device models thus improving robustness. Soliton generation on a nonlinear transmission line is considered as an example. stant time steps can be used further improving the robustness of the convolution. A nonlinear transmission line is used here and it is regarded by many in the field as an extreme test of the performance of transient and steady-state simulators. More generally, the work is directed at the transient simulation of distributed systems with tightly coupled circuit-field interactions. I. Introduction T RANSIENT analysis of distributed microwave circuits is complicated by the inability of frequency independent primitives (such as resistors, inductors and capacitors) to model distributed circuits. More generally, the linear part of a microwave circuit is described in the frequency domain by network parameters especially where numerical field analysis is used to model a spatially distributed structure. Inverse Fourier transformation of these network parameters yields the impulse response of the linear circuit. This has been used with convolution to achieve transient analysis of distributed circuits [1], [2], [3], [4]. The major drawback of convolution analysis has been the large time and memory requirements that result from recording and convolving the nodal voltages of the nonlinear elements. The situation worsens because of convergence considerations which require small time steps. This paper develops a convolution-based transient analysis which uses state variables instead of node voltages to capture the nonlinear response. This has two desirable effects. First, the number of state variables required is less than the number of nodes of the nonlinear elements and so fewer quantities need to be recorded and convolved. Second, parameterized device models can be used as the nonlinearity is not restricted to the form of current as a nonlinear function of voltage. Parameterized device models result in stable transient analysis for larger time steps and so less discretized time history is required. Also with improved stability con- This work was supported by the Defense Advanced Research Projects Agency (DARPA) through the MAFET Thrust III program as DARPA agreement number DAAL01-96-K-3619. C. E. Christoffersen and M. Ozkar are with the Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC 27695-7911, U.S.A., (e-mail: c.christoffersen@ieee.org, mete ozkar@ieee.org). M. B. Steer was with NCSU. He is now with the Institute of Microwaves and Photonics, School of Electronic and Electrical Engineering, University of Leeds, Leeds, United Kingdom LS2 9JT, (e-mail: m.b.steer@ieee.org). M. G. Case was and M. Rodwell is with the department of Electrical and Computer Engineering, University of California, Santa Barbara, CA 93106, U.S.A. II. Background Distributed linear microwave circuits are described by frequency-dependent network parameters and only a few methods are available to accommodate these circuits in transient simulation. These methods include Impulse Response and Convolution (IRC), Asymptotic Waveform Evaluation (AWE), and Laplace Inversion. A. Impulse Response and Convolution One of the first implementations of IRC to distributed microwave circuits was by Djordjevic and Sarkar [5] in 1987. Since that time there have been several efforts to reduce the significant aliasing errors that result from the Inverse Fast Fourier Transform (I-FFT) operation. Minimization of aliasing in the I-FFT requires that the imaginary part of the frequency response be zero at the maximum frequency. Low pass filtering has been used to achieve this [6]. The introduction of a small time delay also achieves this result with presumably less distortion [4]. Insertion of an augmentation network in the linear network at the interface with the nonlinear network achieves the same result for special types of circuits [1]. The effect of the augmentation network is compensated in the nonlinear iteration scheme. It is also important to limit the length of the impulse response to reduce memory requirements and the resistive augmentation achieves this result [1]. Augmentation is also effectively achieved using s parameters [7], where the reference impedance effectively dampens multiple reflections. Distributed networks are characterized partly by reflections so that an impulse response tends to have regions of low value between regions of rapid change. In this case thresholding greatly reduces the number of impulse response discretizations that need to be retained [1]. In high speed digital interconnect circuitry this can reduce the number of significant impulse responses by a factor of 10 to 100, depending on the desired accuracy [1]. Convolution as generally practiced uses a rectangular in- 2 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES tegration scheme (essentially an impulse response is treated as being constant in a time step interval). P. Stenius et al. [8] developed a trapezoidal form of the convolution integration which could possibly have superior convergence properties than the previous block integration. B. Asymptotic Waveform Evaluation The frequency dependent network parameters can be modeled using frequency independent primitives (resistors, inductors and capacitors) if a rational polynomial transfer function is fitted to the network parameters. In practice, this procedure results in an impossibly large circuit. The AWE method addresses this problem by reducing the dimension of the rational polynomial while minimizing distortion [9], [11]. This method works well for interconnects in digital systems and lower frequency microwave circuits [12], [13]. However, higher frequency circuits can only be modeled approximately using AWE due to the infinite number of poles and zeros of such a circuit. Most of the AWE methods use the Padé approximation and this and other approximations used can have stability problems [9]. Recent work has extended this technique to more general distributed circuits [10]. C. Numerical Inversion of Laplace Transform Technique This technique does not have aliasing problems since it does not assume that the function is periodic. The inverse transform exists for both periodic and non-periodic functions. There is no causality problem for double sided Laplace Transforms, either. Unlike FFT methods, the desired part of the response can be achieved without doing tedious and unnecessary calculations for the other parts of the response. Laplace techniques suffer from the series approximations and the nonlinear iterations involved. The advantages and the limitations of the Inverse Laplace Methods are discussed in detail in [14]. III. Formulation of the Transient Analysis As it is now conventional for distributed microwave circuits, the circuit is partitioned into linear and nonlinear subcircuits [15]. The core of the method presented here is that state variables are used to describe the nonlinear dependence in the nonlinear subcircuit. This enables more flexibility in writing the nonlinear element models so that better convergence might be achieved [20]. The linear subcircuit is characterized in the frequency domain and the other part, which includes the nonlinear devices, is treated in the time domain. From [16], the vector of voltages VL in the frequency domain calculated at the interface to the linear network is VL (X, f) = SSV (f) + MSV (f) INL (X, f) (1) where X is the state variable vector, MSV is the state variable impedance matrix and INL is the vector of currents flowing into the linear network at the interface of the linear/nonlinear network. Since the independent sources are more easily handled in the time domain for this kind of analysis, the source vector Ssv (f) can be assumed to be zero and the sources are considered together with the nonlinear devices. Expanding the matrix multiplication, each element of the voltage vector VL (X, f) can be written as: VL i = nS X Mi,j (f) INL,j (2) j=1 Rewriting one term of this equation in the time domain and replacing the multiplications with the convolution operation leads to vLi,j (t) = mi,j (t) ∗ iNLj (t) (3) Now, expanding the convolution operation we get: Z t vL i,j (t) = mi,j (τ ) iNL j (t − τ ) dτ (4) −∞ where the system is assumed to be causal, iNL j (t) = 0 for t ≤ 0. Numerical evaluation requires discretization as follows. First, each element of the I-FFT of MSV has a finite number of components, denoted NT . Using the trapezoidal integration rule [8] and using mi,j obtained from the I-FFT, we obtain Eq. (5). For nt < NT the last term is zero since ij (0) = 0. For nt ≥ NT the last term is also zero since mi,j (nt ) = 0. vLi,j (X̂, nt ) = m (0) i (X̂,n ) i,j NL t 2 P n −1 + t nτ =1 mi,j (nτ ) ij (nt − nτ ), if nt < NT mi,j (0) iNL j (X̂,nt ) PNT 2 + nτ =1 mi,j (nτ ) ij (nt − nτ ), if nt ≥ NT (5) Note that in all but one of the convolution sum terms, iNL = iL = i, so most of the convolution sum can be performed before beginning the iterations to solve the nonlinear system of equations. The following error function is solved at each time step Fi(X̂) = VLi (X̂) − VNL i(X̂) and in vector form, the complete system is Pn s j=1 vL1,j (X̂, nt ) − vNL1 (X̂, nt ) Pn s j=1 vL2,j (X̂, nt ) − vNL2 (X̂, nt ) F(X̂) = .. . Pn s j=1 vLns ,j (X̂, nt ) − vNLns (X̂, nt ) (6) (7) The error function in Eq. (7) is solved for each time step. IV. Implementation The formulation developed above resulted in a standard nonlinear problem that can be solved using the Newton method or quasi-Newton method. As the error function CHRISTOFFERSEN et al.: STATE VARIABLE-BASED TRANSIENT ANALYSIS USING CONVOLUTION changes only slightly from time step to time step, efficient iterative matrix solve schemes can be used as a very good preconditioner is available from the previous time step. The general flow of the analysis is shown in Figure 1 and was implemented in the Transim program. 3 shown in Fig. 2. The advantage of this topology is that no extra nodes are added to the circuit and the size of the MNAM is not changed. The augmentation network provides a better matrix conditioning for the I-FFT operation at the expense of increased error due to finite numerical accuracy. Ideally, the effect of the augmentation would be exactly compensated. The resistive augmentation network Parse Netlist I Li PREPROCESSING VLi VNLi ... ... Create MNAM I NLi ... ... ... ... NONLINEAR ELEMENT Add Augmentation Network Calculate Msv Matrix ... LINEAR SUBCIRCUIT Do Filtering ... ... I-FFT NONLINEAR ELEMENT AT EACH TIME STEP Convolution Sum Solve Nonlinear System AUGMENTATION NETWORK COMPENSATION NETWORK Iterate until convergence Store Currents Output Routines Fig. 1. Flow diagram of the analysis code A. Modified Nodal Admittance Matrix Computation begins with the formulation and solution of a frequency domain modified nodal admittance matrix (MNAM) of the linear subcircuit. The MNAM routines are based on the sparse matrix package Sparse 1.3 [17]. This is a flexible package of subroutines written in C that quickly and accurately solve large sparse systems of linear equations. It also provides utilities such as MNAM reordering and other utilities suited to circuit analysis. At each frequency, the state variable impedance matrix MSV is calculated from the LU decomposed MNAM [16]. B. Impulse Response Determination The impulse response is obtained using the inverse real Fourier transformation of each element of MSV . The transform requires special characteristics of the frequency domain variables at high frequencies so that aliasing is minimized. Problems occur with ideal inductors and capacitors as the matrix elements can go to infinity. This can be corrected by cascading a resistive augmentation circuit with the linear circuit and so ensure finite parameters [18], [1]. The use of scattering parameters achieves similar resistive augmentation [7]. Being resistive, the effect of the augmentation network can be removed in transient simulation as the resistors are unaffected by the Fourier transformation. The current work uses the resistive augmentation network LINEAR / NONLINEAR INTERFACE OF AUGMENTED CIRCUIT Fig. 2. Augmentation and compensation network also serves to limit the extent of the impulse response as energy is absorbed and multiple reflections are damped. Again, note this effect is fully compensated. Before converting the frequency domain MSV (impedancelike) matrix to the time domain, the imaginary part of the frequency response needs to be band-limited so that the function has a periodic (or circular) frequency response. This significantly reduces the aliasing errors in the I-FFT operation. In Transim, frequency response limiting is implemented by filtering the last part of the frequency spectrum of each matrix element so the imaginary part goes to zero and the real part goes to the DC value. In this way, the frequency response is periodic. This is in contrast to low pass filtering [1] which zeroes the high frequency response and introducing additional time-delay to take the imaginary response to zero [2]. The filtering factor ki is, 2 n−i ki = (8) n − i0 where n is the number of discrete frequencies, i is the frequency index and i0 is the frequency index at which filtering starts. For all i greater than i0 , the filtering operation on each state variable impedance matrix entry is <{mi } = ki × <{mi } + (1 − ki ) × m0 ={mi } = ki × ={mi } (9) (10) The amount of filtering is an analysis option since in some cases it is not necessary. By default, only the last 3% of IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES the spectrum is filtered to achieve an acceptable alias reduction. The effect on causality remains to be determined. If this is not done the correct impulse response is not assured as the continuity condition is inherent in the I-FFT (or FFT) operation. Causality imposes special considerations for the response obtained from the I-FFT operation. A discontinuity of the impulse response generally occurs at t = 0, as the negative time response must be zero [8]. In this situation the I-FFT operation yields a value which is the average of the one-sided limits of the function at that point. That is the value calculated at t = 0 is one half of the value at t = 0+ . However, the upper side limit (at t = 0+ ) is required in numerical integration and so the value of the I-FFT calculated time response must be multiplied by 2 to obtain the required impulse response. Note that in Equation (5) the first point of the impulse is divided by two, so actually the multiplication can be saved. Fourier transformation is implemented using a real IFFT algorithm from the FFTW library [19]. This algorithm requires only the positive frequency samples since the negative frequency samples are the complex conjugate of the corresponding positive frequencies. The FFTW library is a comprehensive collection of C routines for computing the discrete Fourier transform in one or more dimensions, of both real and complex data, and of arbitrary input size. C. Parameterized Nonlinear Models Let the nonlinear subnetwork be described by the following generalized parametric equations [20] vNL (t) iNL (t) dx dn x , . . . , n , xD (t)] dt dt dx dn x = w[x(t), , . . . , n , xD (t)] dt dt = u[x(t), x(t) if x(t) ≤ V1 (14) 1 V + ln(1 + α(x(t) − V1 )) 1 α if x(t) > V1 Is (exp(αx(t)) − 1) if x(t) ≤ V1 = (15) I exp(αV )(1 + α(x(t) − V1 )) − Is s 1 if x(t) > V1 v(t) = i(t) where V1 is some threshold value. The model requires the current, voltage and derivatives to be continuous at x = V1 . This implies that V1 = ln(G1 /αIs ) α (16) where G is the slope ∂i/∂v and is chosen to be 5e8 · Is . Note that using this value, V1 becomes independent of the saturation current. In this way, the maximum value of the exponential function is limited to 5e8. The plots of Figures 3, 4 and 5 show the improvement in the behavior of the model when using the state variable approach. In a diode the current has an exponential dependence on voltage. This causes convergence problems when the voltage is updated during nonlinear iteration. At voltages greater than the threshold, small voltage increments can result in large current changes and hence changes in the error function. The possibility of large changes is eliminated through the use of parameterization which ensures smooth, well behaved current, voltage and error function variations when the state variable is updated. See Figures 4 and 5. (11) 0.6 (12) 0.5 where vNL (t), iNL (t) are vectors of voltages and currents at the common ports, x(t) is a vector of state variables and xD (t) a vector of time-delayed state variables, i.e., xDi(t) = xi (t − τi). The time delays τi may be functions of the state variables. All vectors in (12) have the same size nd equal to the number of common (device) ports. This kind of representation is convenient from the physical viewpoint, as it is equivalent to a set of implicit integro-differential equations in the port currents and voltages. This allows an effective minimization of the number of subnetwork ports, and what is more important, results in extreme generality in device modeling capabilities. Here a parametric diode model for the diode element is presented [20]. The full model includes capacitances and nonlinear resistance, but here a simple case is shown for didactic purposes. The conventional current equation for the diode is i(t) = Is (exp(αv(t)) − 1) as follows (13) and, based on previous work [20], the parametric model is 0.4 0.3 I (A) 4 0.2 0.1 0 −0.1 −1 −0.8 −0.6 −0.4 −0.2 0 V (V) 0.2 0.4 0.6 0.8 1 Fig. 3. Relation between v and i in a diode. D. Convolution Convolution of the impulse response of the linear circuit with the outputs of the nonlinear elements has been proposed for transient analysis of distributed circuits with [7], CHRISTOFFERSEN et al.: STATE VARIABLE-BASED TRANSIENT ANALYSIS USING CONVOLUTION 0.6 5 vides Newton and quasi-Newton methods and many options, such as the use of analytic Jacobian or forward, backwards or central differences to approximate it, different quasi-Newton Jacobian updates and two globally convergent methods. This flexibility permits rapid simulator development. Additionally, it reflects state-of-the-art numerical analysis. 0.5 0.4 I (A) 0.3 V. Simulation Of A Nonlinear Transmission Line 0.2 0.1 0 −0.1 −1 −0.5 0 0.5 1 1.5 2 2.5 X Fig. 4. Relation between x and i in a diode. 1 0.8 0.6 0.4 V (V) 0.2 0 −0.2 −0.4 Nonlinear transmission lines (NLTLs) find applications in a variety of high speed, wide bandwidth systems including picosecond resolution sampling circuits, laser and switching diode drivers, test waveform generators, and mmwave sources [23]. They have three fundamental characteristics: nonlinearity, dispersion and dissipation. The NLTL consist of coplanar waveguides (CPWs) periodically loaded with reverse biased Schottky diodes. Diode-based NLTL used for pulse generation are extremely nonlinear circuits and are being used to test the robustness of circuit simulators. The NLTL considered here was designed with a balance between the nonlinearity of the loaded nonlinear elements and the dispersion of the periodic structure which results in the formation of a stable soliton [24], [25]. The nonlinearity of NLTLs is principally due to the voltage dependent capacitance of the diodes and the dissipation is due to the conductor losses in the CPWs. In this work, the NLTL was modeled using generic transmission lines with frequency-dependent loss and Schottky diodes [26]. Skin effect was taken into account in the modeling of the transmission lines. The NLTL model is shown in Figure 6 and is excited by a 9 GHz sinusoid. The NLTL −0.6 50Ω + −0.8 −1 −1 −0.5 0 0.5 1 1.5 2 0011 0011 ... 0011 50Ω 2.5 X Fig. 5. Relation between x and v in a diode. [2], [3], [4], [6], [1], [21]. The major problem identified with this type of analysis is the rapid growth in computation and memory requirements. The convolution integral, which becomes a convolution sum in computer implementations, is O(n2M AX nS 2 ) where nM AX is the length of the impulse response and nS is the number of state variables. Memory usage is O(nM AX × nS 2 ), principally due to storage of the matrix impulse response. Thus reducing the number of state variables, nS , (i.e., using the minimum number of state variables rather than the voltages at all nodes of the nonlinear network) dramatically reduces memory and compute time. As state variables permit the use of parameterized device models, the stability of the convolution analysis is improved allowing larger time steps thus reducing nM AX . E. Nonlinear Equation Solver The nonlinear system at each time step is solved using the NNES library [22]. It is written in Fortran and pro- Fig. 6. Model of the nonlinear transmission line. was designed for a 24 GHz initial Bragg frequency, 225 GHz final Bragg frequency, 0.952097 tapering rule, and 120 ps total compression. It contains 48 sections of CPW transmission lines and 47 diodes. The drive is a 27 dBm sine wave with −3 V dc bias. VI. Results and Discussion Fig. 7 shows the calculated transient response of the soliton line including frequency dependent (skin effect) loss of the transmission lines. The simulation time was 8 hours on an UltraSPARC 1 workstation, clocked at 143 MHz. The number of frequencies used to obtain the impulse response was 4096 and the maximum frequency was 2700 GHz, which corresponds to a time step of 0.185ps. The value of the compensation resistor was 140 Ω. The comparison among the measured output voltage across the load [23] and transient and harmonic balance simulation (also using Transim) is shown in Fig. 8. The harmonic balance analysis used the state variable approach described in [16]. The convergence rate was increased and the number of harmonics reduced using the filtering technique described in [27], 6 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES 30 0 25 −2 20 15 OUTPUT POWER (dBm) VOLTAGE (V) −4 −6 −8 −10 10 5 0 −5 −10 −12 −15 −14 0 100 200 300 400 500 TIME (ps) 600 700 800 900 −20 1000 0 50 100 150 200 SPECTRUM (GHz) 250 300 350 Fig. 7. Complete transient response for the soliton line Fig. 9. Magnitude of the harmonics of the output power. starting at harmonic number 30. The total number of frequencies was 40, and the simulation time was 30 hours on the same computer. The harmonic content of the output is shown in Fig. 9. The prediction of the main soliton am0 continues on identify the required time step for a prescribed accuracy. The importance of IRC analysis of distributed microwave circuits as opposed to conventional SPICE analysis is that frequency dependent linear elements can be handled. In Fig. 11 the effect of frequency dependent trans- −5 −2 −4 −10 VOLTAGE (V) OUTPUT VOLTAGE (V) 0 EXPERIMENTAL DATA HARMONIC BALANCE TRANSIENT −6 −8 −10 −15 0 10 20 30 TIME (ps) 40 50 60 −12 .18ps .13ps .27ps Fig. 8. Comparison between experimental data and simulations −14 380 plitude is correct, although the impulse width is slightly smaller in the simulations. Note that the secondary soliton predicted by the simulations is not observed in the measurements. This is probably due to neglecting higher order parasitic effects of the interconnections. In both simulations, a parasitic inductance of 21.8 pH modeled the connection between each diode and the CPW line. Without adding complexity this inductor was incorporated in the nonlinear diode model. This is only possible with state variables and results in a better conditioned transient analysis. The transient analysis gives slightly different responses depending on the size of the time step as shown in Fig. 10. Smaller time step yields a simulation more accurate. Work 400 420 440 TIME (ps) 460 480 500 Fig. 10. Comparison between simulations using different time steps mission line attenuation, principally due to the skin effect, is compared to frequency-independent attenuation and no attenuation. The frequency independent attenuation of the transmission lines is the attenuation at 10 GHz. Frequency dependent attenuation has only a small effect on the depth and width of the soliton. The importance of modeling is borne out by these results. Rather than focusing on the previously deleterious effect of frequency dependent loss, better modeling of the parasitics of the NLTL is required to provide a better basis for computer aided design. CHRISTOFFERSEN et al.: STATE VARIABLE-BASED TRANSIENT ANALYSIS USING CONVOLUTION 0 [7] −2 [8] VOLTAGE (V) −4 −6 [9] −8 [10] −10 [11] −12 ACTUAL FREQUENCY−INDEPENDENT NONE −14 830 840 850 860 870 880 TIME (ps) 890 900 910 920 Fig. 11. Comparison between simulations using different attenuation factors for the transmission lines [12] [13] [14] VII. Conclusion The importance of the work presented here was the development of a state-variable formulation for transient analysis of distributed circuits using the impulse response of the linear subnetwork and convolution. State variable based analysis minimizes compute time and memory requirements when the number of linear elements with frequency-dependent loses is greater or about the same as the number of nonlinear elements, but most importantly it improves the robustness of the simulation by allowing parameterized nonlinear device models to be used. The transient simulation technique developed here is targeted at the analysis of circuits with tightly coupled circuit and field interactions. Modeling of the electromagnetic environment results in port descriptions without a defined global reference node rendering the use of nodal voltages is problematic [28], [29]. The current implementation of Transim circumvents this problem by the use of local reference nodes. [15] [16] [17] [18] [19] [20] [21] References [1] [2] [3] [4] [5] [6] M. S. Basel, M. B. Steer and P. D. Franzon, “Simulation of high speed interconnects using a convolution-based hierarchical packaging simulator,” IEEE Trans. on Components, Packaging, and Manufacturing Techn., Vol. 18, February 1995, pp. 74-82. T. J. Brazil, “A new method for the transient simulation of causal linear systems described in the frequency domain,” 1992 IEEE MTT-S Int. Microwave Symp. Digest, June 1992, pp. 1485-1488. P. Perry and T. J. Brazil, “Hilbert-transform-derived relative group delay,” IEEE Trans. on Microwave Theory and Techn., Vol 45, Aug. 1997, pt. 1, pp. 1214-1225. T. J. Brazil, “Causal convolution—a new method for the transient analysis of linear systems at microwave frequencies,” IEEE Trans. on Microwave Theory and Techn., Vol. 43, Feb. 1995, pp. 315-23. A. R. Djordjevic and T. K. Sarkar, “Analysis of time response of lossy multiconductor transmission line networks,” IEEE Trans. on Microwave Theory and Techn., Vol. MTT-35, Oct. 1987, pp. 898-908. D. Winkelstein, R. Pomerleau and M. B. Steer, “Transient simulation of complex, lossy, multi-port transmission line networks with nonlinear digital device termination using a circuit simu- [22] [23] [24] [25] [26] [27] 7 lator,” Conf. Proc. IEEE SOUTHEASTCON, Vol. 3, pp. 12391244. J. E. Schutt-Aine and R. Mittra, “Nonlinear transient analysis of coupled transmission lines,” IEEE Trans. on Circuits and Systems, Vol. 36, Jul. 1989, pp. 959-967. P. Stenius, P. Heikkilä and M. Valtonen, “Transient analysis of circuits including frequency-dependent components using transgyrator and convolution,” Proc. of the 11th European Conference on Circuit Theory and Design, Part II, 1993, pp. 1299-1304. P. K. Chan, Comments on “Asymptotic waveform evaluation for timing analysis,” IEEE Trans. on Computer Aided Design, Vol. 10, Aug. 1991, pp. 1078-79. R. Archar, M. S. Nakhla and Qi-Jun Zhang, “Full-wave analysis of high speed interconnects using complex frequency hoping,” IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems, Oct. 1998, pp 997-1010. M. Celik, O. Ocali, M. A. Tan, and A. Atalar, “Pole-zero computation in microwave circuits using multipoint Padé approximation,” IEEE Trans. on Circuits and Systems, Jan. 1995, pp. 6-13. E. Chiprout and M. Nakhla, “Fast nonlinear waveform estimation for large distributed networks,” 1992 IEEE MTT-S Int. Microwave Symp. Digest, Vol.3, Jun. 1992, pp. 1341-1344. R. J. Trihy and Ronald A. Rohrer, “AWE macromodels for nonlinear circuits,” Proceedings of the 36th Midwest Symposium on Circuits and Systems, Vol. 1, Aug. 1993, pp. 633-636. R. Griffith and M. S. Nakhla, “Mixed frequency/time domain analysis of nonlinear circuits,” IEEE Trans. on Computer Aided Design, Vol.11, Aug. 1992, pp. 1032-43. M. S. Nakhla and J. Vlach, “A piecewise harmonic balance technique for determination of periodic response of nonlinear systems,” IEEE Trans. Circuits Systems, Vol. 23, Feb. 1976, pp. 85-91. C. E. Christoffersen, M. B. Steer and M. A. Summers, “Harmonic balance analysis for systems with circuit-field interactions,” 1998 IEEE Int. Microwave Symp. Digest, June 1998, pp. 1131-1134. K. S. Kundert and A. Songiovanni-Vincentelli, Sparse user’s guide - a sparse linear equation solver, Dept. of Electrical Engineering and Computer Sciences, University of California, Berkeley, Calif. 94720, Version 1.3a, Apr 1988. M. Ozkar, Transient analysis of spatially distributed microwave circuits using convolution and state variables, M. S. Thesis, Department of Electrical and Computer Engineering, North Carolina State University. M. Frigo and S. G. Johnson, FFTW User’s Manual, Massachusetts Institute of Technology, September 1998. V. Rizzoli, A. Lipparini, A. Costanzo, F. Mastri, C. Ceccetti, A. Neri and D. Masotti, “State-of-the-art harmonic-balance simulation of forced nonlinear microwave circuits by the piecewise technique,” IEEE Trans. on Microwave Theory and Techn., Vol. 40, Jan 1992, pp 12-27. C. Gordon, T. Blazeck and R. Mittra, “Time domain simulation of multiconductor transmission lines with frequency-dependent losses,” IEEE Trans. on Computer Aided Design of Integrated Circuits and Systems, Vol. 11 Nov. 1992 pp. 1372-87. R. S. Bain, NNES user’s manual, 1993. M. G. Case, Nonlinear transmission lines for picosecond pulse, impulse and millimeter-wave harmonic generation, Ph.D Dissertation, Department of Electrical and Computer Engineering, University of California, Santa Barbara, California, U.S.A., 1993. M. J. W. Rodwell, M. Kamegawa, R. Yu, M. Case, E. Carman and K. S. Giboney, “GaAs nonlinear transmission lines for picosecond pulse generation and millimeter-wave sampling,” IEEE Trans. on Microwave Theory and Techn., Vol. 39, July 1991, pp. 1194-1204. H. Shi, C. W. Domier, N. C. Luhmann, “A monolithic nonlinear transmission line system for the experimental study of lattice solutions,” J. of Applied Physics, Vol 4., August 1995, pp. 255864. Compact Software, Microwave Harmonica Elements Library, 1994. A. Brambilla and D. D’Amore, “A filter-based technique for the harmonic balance method,” IEEE Trans. on Circuits and Systems—I: Fundamental Theory and Applications, Vol. 43, Feb. 1996, pp. 92-98. 8 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES [28] A. I. Khalil and M. B. Steer “Circuit theory for spatially distributed microwave circuits,” IEEE Trans. on Microwave Theory and Techn., Vol. 46, Oct. 1998, pp 1500-1503. [29] C. E. Christoffersen and M. B. Steer “Implementation of the local reference concept for spatially distributed circuits,” Int. J. of RF and Microwave Computer-Aided Eng., In Press. Carlos E. Christoffersen (S’91) was born in Argentina in 1968. He received the Electronic Engineer degree at the National University of Rosario, Argentina in 1993. From 1993 to 1995 he was research fellow of the National Research Council of Argentina (CONICET). He received the M.S. degree in electrical engineering in 1998 from North Carolina State University. Currently he is pursuing his Ph.D. degree at the same university. Since 1996, he is with the Department of Electrical and Computer Engineering as a research assistant. His research interests include computer aided analysis of circuits, analog, RF and microwave circuit design. Mete Ozkar (S’97) received the B.S. degree in electrical engineering from Middle East Technical University, Ankara, Turkey, in 1996, and the M.S. degree from North Carolina State University, Raleigh, North Carolina, in 1998. He is currently working toward the Ph.D degree in electrical engineering and is a Research Assistant with the Electronics Research Laboratory, Department of Electrical and Computer engineering, North Carolina State University, where he performs research in the simulation and the development of spatial power combining systems. Michael Case (S’87-M’93-SM’98) was born in 1966 in Ventura, California. After receiving his B. S. degree in 1989 from U. C. Santa Barbara, he began researching there under a state fellowship for Professor Mark Rodwell. In Dr. Rodwell’s group he studied nonlinear transmission lines for applications in high-speed waveform shaping and signal detection. Michael earned his M. S. and Ph. D. degrees in 1991 and 1993 respectively from U. C. Santa Barbara. The Air Force Office of Scientific Research sponsored a majority of his work. HRL Laboratories (previously the Hughes Aircraft Company’s research labs) has employed Michael in Malibu, California since 1993. He is currently involved with millimeter-wave device characterization, circuit design, and measurement techniques. Mark Rodwell was born in Altrincham, England in 1960. He received the B.S. degree in electrical engineering from the University of Tennessee, Knoxville, in 1980, and the M.S. (1982) and Ph.D. degrees (1988) in electrical engineering from Stanford University. From 1982 through 1984 he worked at AT&T Bell Laboratories, developing optical transmission systems. He was a research associate at Stanford University from January to September 1988. In September 1988 he joined the Department of Electrical and Computer Engineering, at the University of California, Santa Barbara, where he is currently Professor and Director of the Compound Semiconductor Research Laboratories. His current research involves submicron scaling of millimeter-wave heterojunction bipolar transistors (HBTs) and development of HBT integrated circuits for microwave mixed-signal ICs and fiber optic transmission systems. His group has developed deep submicron Schottky-collector resonant-tunnel diodes with THz bandwidths, and has developed monolithic submillimeter-wave oscillators with these devices. His group has worked extensively in the area of GaAs Schottky-diode integrated circuits for subpicosecond pulse generation, signal sampling at submillimeter-wave bandwidths, and millimeter-wave instrumentation. He is the recipient of a 1989 National Science Foundation Presidential Young Investigator award, and his work on submillimeterwave diode ICs was awarded the 1997 IEEE Microwave Prize . Michael B. Steer (S’76—M’78—SM’90—F’99) for a photograph and biography, see p. 13 of the January 1999 issue of this Transactions.